Title: Utilizing Nonnegative Tensor Factorization Methods for Inference, Model Selection, and Analysis in Supervised Learning
Advisor(s): A. Taylan Cemgil, Arzucan Özgür
Abstract: This thesis focuses on utilizing nonnegative tensor factorization methods in various areas of supervised learning. The thesis starts with the introduction of a latent variable modeling framework that generalizes different probabilistic approaches to analyzing relational data, including but not limited to nonnegative tensor factorization methods. This framework allows convenient inference and model selection through variational and sampling-based approaches for the model families it encompasses. The flexibility provided by this methodology is then utilized for inference, model selection, and analysis in various supervised learning problems. First, in addition to modeling relational data, we use this framework to model time series with multiple seasonalities. We then demonstrate the use of this framework for learning when to defer to experts in a classification problem, based on the output of one or more machine learning models. Afterwards, we use the framework to investigate the factors that contribute to robust generalization in different supervised learning scenarios. Our work shows that access to a convenient probabilistic framework for modeling relational data can be beneficial for a wide selection of supervised learning problems.
